An open, serverless platform demonstrating real-world AI for contact centers — voice bots, IVR automation, real-time agent assist, NLP pipelines and telephony integration. Built entirely on AWS free-tier.
Each component is independently deployable, production-pattern quality, and runs near-free on AWS serverless infrastructure.
Replace rigid DTMF menu trees with conversational NLU. Customers speak naturally — the system detects intent, fills required slots, executes backend actions via Lambda, and only escalates when necessary.
Multi-turn stateful dialogue with context retention across turns, pronoun resolution, and warm transfer with full context to human agents via SIP UUI.
Live transcription feeds NLP analysis. Next-best-action, KB article suggestions, compliance alerts and sentiment scoring — delivered to the agent desktop within 300ms.
Fine-tuned transformer pipeline for intent classification, NER (account numbers, dates, amounts) and confidence-threshold fallback routing with zero-shot coverage.
SIP-based integration connecting AI voice bots to enterprise telephony — Genesys, Avaya, Five9. SIPREC call recording, MRCP STT/TTS, CTI screen-pop automation.
Live web fetch RAG using BAAI/bge-small-en-v1.5 embeddings. Cosine similarity retrieval, warm-container semantic cache at threshold 0.97, parallel fetch across 15 threads.
Every request flows through a five-stage serverless pipeline — no orchestration servers, no queues, no idle cost.
Inbound call arrives via SIP trunk into Amazon Connect, Genesys Cloud, or Five9. Audio is streamed in real-time via WebSocket at 8kHz μ-law PCM. DTMF fallback is always present for accessibility compliance.
Audio chunks stream to AWS Transcribe via WebSocket. Partial transcripts return in real-time — voice activity detection (VAD) fires end-of-utterance within 120ms of silence, cutting latency by 200–400ms vs. fixed timeouts.
Transcript feeds the NLU Lambda. BAAI/bge-small-en-v1.5 embeddings via FastEmbed, fine-tuned RoBERTa for intent classification, custom NER for account numbers, dates and amounts. Semantic cache checked first — cache hit returns in under 10ms.
Dialogue manager maintains session state in Redis (keyed by call-session-ID). Slot filling collects missing entities. Lambda calls CRM REST API, order systems or knowledge base. If confidence < 0.70, clarification is triggered. Two failed turns → escalate to agent with full context.
Response generates via Amazon Polly Neural TTS — first audio chunk in ~80ms via streaming. If escalating to an agent: session serialised, SIP REFER with UUI context token, CRM case pre-created, agent whisper message played before connection. Total p95 round-trip under 400ms.
Eight purpose-built capabilities covering the full stack of modern contact center intelligence.
Real-time audio streaming via WebSocket. Partial transcript feedback. VAD-triggered end-of-speech. Supports 8kHz telephony and 16kHz wideband.
AWS Transcribe · DeepgramFine-tuned transformer models on contact center transcripts. 35+ intent classes. Named entity extraction for financial, temporal and account entities.
RoBERTa · BART-MNLI · FastEmbedStateful session management with context carryover across turns. Pronoun resolution, slot filling forms, interruption handling and graceful recovery.
Redis · Dialogflow CX · Custom FSMLive web fetch RAG with cosine similarity ranking. Warm-container semantic cache at 0.97 threshold. Parallel multi-thread fetch for sub-second response.
BAAI/bge-small · NumPy · Lambda LayerSIP trunk connectivity, SIPREC media forking, MRCP ASR/TTS resource control, CTI middleware for agent desktop screen-pop automation.
SIP · SIPREC · MRCP · CTITranscript batch processing with BERTopic modelling, sentiment trending, escalation pattern detection and VOC dashboard generation.
Kinesis · Athena · QuickSightLambda-backed login with hashed credential verification. Session token stored client-side. Every page guards on load with auto-redirect. Zero server maintenance.
AWS Lambda · API Gateway · localStorageUpload PDF, TXT or HTML documents and query their content via semantic chunking and keyword scoring. Lambda-backed analysis returns best-matching sections.
Client-side chunking · Lambda RAGThe entire platform runs on five AWS services — fully serverless, independently scalable, pennies per thousand requests.
Pre-built integration patterns for all major CCaaS platforms, CRM systems and telephony infrastructure.
The future of contact centers isn't cheaper agents — it's intelligent systems that resolve what matters before a human ever needs to answer.
— Viswanext · Voice AI Platform Initiative
Sign in to explore the full Voice AI dashboard — live demo chat, architecture diagrams, project showcase, and document intelligence — all in one place.
Authentication via AWS Lambda · Session-based access · Viswanext Initiative